An Earthquake Early Warning (EEW) system generates a warning of forthcoming hazardous part of the strong ground motion to prevent failure of safety-critical infrastructures and reduce death and injuries. The advent of state-of-the-art technologies has inspired many researchers and authorities to build reliable and robust EEW networks. Two significant reference parameters for EEW are seismic intensity and Peak Ground Acceleration (PGA). However, the ground motion estimation problem is a complex task due to the high variability of the ground medium compared to the availability of seismic data. Despite the complexity, machine learning has been used for seismicity prediction, magnitude estimation, magnitude forecasting, and EEW. Hence, in this thesis, early prediction of PGA and estimation of earthquake intensity is attempted using Multilayer Perceptron (MLP). Earthquake data collected from K-Net, Japan and nonearthquake data generated in a span of four years by five Seismic Sensing Nodes (SSN) placed around NCR, India are used. As a regression analysis, PGA is predicted from features extracted from consecutive one-second feature windows. The logarithmic transformation of PGA improved the regression performance. The feature window of 5-6 seconds provides the best r2 score, MAE and MSE of 77.74, 0.35, and 0.19 respectively. PGA values are converted into severity categories by intensitybased MMI scale. Being rare natural events, limited data related to strong earthquakes is available. This warning class imbalance is overcome using stratified differential resampling, which increased the mean accuracy of the MLP classifier from 0.76 to 0.91. The f1-score also jumped from 0.79 to 0.91 in the case of the balanced dataset. The use of the genetic algorithm for training neuroevolutionary MLP neural networks reduced the size of the required training dataset to 20% only. Using 5-fold inverse cross-validation, specificity and precision of 95.03% and 94.47% are achieved in this comparatively large test set comprising 80% samples. Further, Faster Than Real Time (FTRT) simulation approach is adopted for assessing online warning prediction performance. A novel Factor of Early Warning (FoEW) is introduced to measure the timeliness of warning for true positive cases. Using the onsite estimates by MLP-GANN, a centralised EEW framework is proposed. The centralised warning is raised at the instance when MLP-GANN models of three stations predict ensuing strong intensity at corresponding sites. The performance of the centralised neuroevolutionary warning framework is validated at sites situated at 3 different radii from the centroid location (ܮ (௚௖of the first three triggered stations. By regression analysis a dynamic radius ܦௐ is calculated for precise warning generation. The accuracy of warning improved from 91% to 95% in the centralised framework. A distribution of Warning Lead Time and FoEW is determined for different Magnitude and source-to-site distances. The simulations yielded more than 85% mean FoEW from PGA arrival, for events with magnitude 5 and above, in all the 3 approaches of warning radii. At sites more than 100 km from the epicentre, more than 25 s of Lead Time is achieved